The field of adversarial robustness has long established that adversarial examples can successfully transfer between image classifiers and that text jailbreaks can successfully transfer between language models (LMs). However, a pair of recent studies reported being unable to successfully transfer image jailbreaks between vision-language models (VLMs). To explain this striking difference, we propose a fundamental distinction regarding the transferability of attacks against machine learning models: attacks in the input data-space can transfer, whereas attacks in model representation space do not, at least not without geometric alignment of representations. We then provide theoretical and empirical evidence of this hypothesis in four different settings. First, we mathematically prove this distinction in a simple setting where two networks compute the same input-output map but via different representations. Second, we construct representation-space attacks against image classifiers that are as successful as well-known data-space attacks, but fail to transfer. Third, we construct representation-space attacks against LMs that successfully jailbreak the attacked models but again fail to transfer. Fourth, we construct data-space attacks against VLMs that successfully transfer to new VLMs, and we show that representation space attacks can transfer when VLMs' latent geometries are sufficiently aligned in post-projector space. Our work reveals that adversarial transfer is not an inherent property of all attacks but contingent on their operational domain - the shared data-space versus models' unique representation spaces - a critical insight for building more robust models.

Recent breakthroughs have spurred claims that large language models (LLMs) match gold medal Olympiad to graduate level proficiency on mathematics benchmarks. In this work, we examine these claims in detail and assess the extent to which current benchmarks capture genuine LLM mathematical reasoning. The composition of these benchmarks, primarily drawing from the International Mathematics Olympiad (IMO) and related competitions, may overstate models reasoning ability due to potential data contamination and a narrow focus on familiar problem types. To enable a more holistic assessment of mathematical understanding, we introduce EEFSUVA, a novel benchmark curated from under circulated regional and national Olympiads of Eastern Europe and the countries from the former Soviet Union. These contests feature problems of comparable difficulty to the IMO and are renowned for demanding nonstandard problem-solving techniques, yet their problems are far less prevalent in online corpora. Preliminary results suggest that even state-of-the-art LLMs exhibit a notable performance decline on EEFSUVA relative to other Olympiad-style benchmarks. These findings also suggest the potential importance of broader evaluation datasets for a fuller assessment of mathematical reasoning and for guiding future model development.

In this work, we reconceptualize autonomous driving as a generalized language and formulate the trajectory planning task as next waypoint prediction. Our framework presents a single-pass generation paradigm that aligns with the inherent sequentiality of driving. This approach leverages the generative capacity of the VLM (Vision-Language Model) to enable end-to-end trajectory prediction directly from front-view camera input. The efficacy of this method is underpinned by a principled supervision strategy derived from statistical modeling. This provides a well-defined learning objective, which makes the framework highly amenable to master complex driving policies through imitation learning from large-scale expert demonstrations. Empirically, our method achieves the state-of-the-art performance on the nuScenes dataset, delivers an overall improvement of over 30% compared to prior baselines. Furthermore, it exhibits superior generalization performance on cross-domain datasets acquired from diverse vehicles, demonstrating notable potential for cross-vehicle robustness and adaptability. Due to these empirical strengths, this work introduces a model enabling fundamental driving behaviors, laying the foundation for the development of more capable self-driving agents. Code will be available upon publication.Figure 1: Visualization of typical driving scenarios. Predicted trajectories and ego vehicle coverage are shown in green, whereas ground truth trajectories are displayed in orange. Human driving is an inherently sequential decision-making process, in which each action is conditioned on a real-time understanding of the surrounding scene. This dynamic interplay of perception and action exhibits strong similarities to natural language generation, which also involves producing a highly correlated sequence of outputs. Viewing the driving task from this perspective allows us to frame a Vision-Language Model (VLM) as a powerful policy network.

Large Language Model (LLM) safety is one of the most pressing challenges for enabling wide-scale deployment. While most studies and global discussions focus on generic harms, such as models assisting users in harming themselves or others, enterprises face a more fundamental concern: whether LLM-based agents are safe for their intended use case. To address this, we introduce operational safety, defined as an LLM's ability to appropriately accept or refuse user queries when tasked with a specific purpose. We further propose OffTopicEval, an evaluation suite and benchmark for measuring operational safety both in general and within specific agentic use cases. Our evaluations on six model families comprising 20 open-weight LLMs reveal that while performance varies across models, all of them remain highly operationally unsafe. Even the strongest models - Qwen-3 (235B) with 77.77% and Mistral (24B) with 79.96% - fall far short of reliable operational safety, while GPT models plateau in the 62-73% range, Phi achieves only mid-level scores (48-70%), and Gemma and Llama-3 collapse to 39.53% and 23.84%, respectively. While operational safety is a core model alignment issue, to suppress these failures, we propose prompt-based steering methods: query grounding (Q-ground) and system-prompt grounding (P-ground), which substantially improve OOD refusal. Q-ground provides consistent gains of up to 23%, while P-ground delivers even larger boosts, raising Llama-3.3 (70B) by 41% and Qwen-3 (30B) by 27%. These results highlight both the urgent need for operational safety interventions and the promise of prompt-based steering as a first step toward more reliable LLM-based agents.

In this paper we summarize the results of the Putnam-like benchmark published by Google DeepMind. This dataset consists of 96 original problems in the spirit of the Putnam Competition and 576 solutions of LLMs. We analyse the performance of models on this set of problems to verify their ability to solve problems from mathematical contests.

Steering has emerged as a promising approach in controlling large language models (LLMs) without modifying model parameters. However, most existing steering methods rely on large-scale datasets to learn clear behavioral information, which limits their applicability in many real-world scenarios. The steering vectors extracted from small dataset often contain task-irrelevant noising features, which degrades their effectiveness. To refine the steering vectors learned from limited data, we introduce Refinement of Steering V ector via Sparse Autoencoder (SAE-RSV) that leverages SAEs to semantically denoise and augment the steering vectors. In our framework, we first remove task-irrelevant features according to their semantics provided by SAEs, and then enrich task-relevant features missing from the small dataset through their semantic similarity to the identified relevant features. Extensive experiments demonstrate that the proposed SAE-RSV substantially outperforms all the baseline methods including supervised fine-tuning. Our findings show that effective steering vector can be constructed from limited training data by refining the original steering vector through SAEs. Large language models (LLMs) have demonstrated remarkable capabilities across a wide range of natural language processing tasks.

Post-training of large language models involves a fundamental trade-off between supervised fine-tuning (SFT), which efficiently mimics demonstrations but tends to memorize, and reinforcement learning (RL), which achieves better generalization at higher computational cost. Dynamic Fine-Tuning (DFT) recently emerged as a promising middle ground, reweighting SFT objectives with token probabilities and achieving improvements in certain reasoning domains, though it exhibits instability in other tasks. We provide a analysis of DFT through the reward-weighted regression (RWR) framework, revealing that it corresponds to a specific auxiliary distribution choice that yields provably tighter RL bounds than standard SFT. However, our analysis also uncovers a critical limitation: this construction lacks distributional anchoring, leading to progressive drift that undermines training stability. To address this, we propose Anchored Supervised Fine-T uning (ASFT), which augments DFT's reweighting with lightweight KL regularization to preserve tightness while ensuring stability. Empirically, ASFT consistently outperforms both SFT and DFT across mathematical reasoning, medical knowledge grounding, and code generation, achieving substantial improvements with minimal computational overhead. Our RWR framework provides a systematic lens for understanding post-training methods and demonstrates that principled theoretical analysis leads to both stronger guarantees and practical gains. Large language models (LLMs) have become a central substrate for modern AI systems, powering instruction following, tool use, and multi-step reasoning at scale (Brown et al., 2020; Touvron et al., 2023; Achiam et al., 2023; Guo et al., 2025).

The evolution toward 6G networks is being accelerated by the Open Radio Access Network (O-RAN) paradigm -- an open, interoperable architecture that enables intelligent, modular applications across public telecom and private enterprise domains. While this openness creates unprecedented opportunities for innovation, it also expands the attack surface, demanding resilient, low-cost, and autonomous security solutions. Legacy defenses remain largely reactive, labor-intensive, and inadequate for the scale and complexity of next-generation systems. Current O-RAN applications focus mainly on network optimization or passive threat detection, with limited capability for closed-loop, automated response. To address this critical gap, we present an agentic AI framework for fully automated, end-to-end threat mitigation in 6G O-RAN environments. MobiLLM orchestrates security workflows through a modular multi-agent system powered by Large Language Models (LLMs). The framework features a Threat Analysis Agent for real-time data triage, a Threat Classification Agent that uses Retrieval-Augmented Generation (RAG) to map anomalies to specific countermeasures, and a Threat Response Agent that safely operationalizes mitigation actions via O-RAN control interfaces. Grounded in trusted knowledge bases such as the MITRE FiGHT framework and 3GPP specifications, and equipped with robust safety guardrails, MobiLLM provides a blueprint for trustworthy AI-driven network security. Initial evaluations demonstrate that MobiLLM can effectively identify and orchestrate complex mitigation strategies, significantly reducing response latency and showcasing the feasibility of autonomous security operations in 6G.

This paper introduces a novel approach to position embeddings in transformer models, named "Exact Positional Embeddings" (ExPE). An absolute positional embedding method that can extrapolate to sequences of lengths longer than the ones it was trained on. Traditional transformer models rely on absolute or relative position embeddings to incorporate positional information into token embeddings, which often struggle with extrapolation to sequences longer than those seen during training. Our proposed method utilizes a novel embedding strategy that encodes exact positional information by overriding specific dimensions of the embedding vectors, thereby enabling a more precise representation of token positions. The proposed approach not only maintains the integrity of the original embeddings but also enhances the model's ability to generalize to more extended sequences. In causal language modeling, our ExPE embeddings significantly reduce perplexity compared to rotary and sinusoidal embeddings, when tested on sequences longer than those used in training.

Large Language Models (LLMs) have made remarkable progress in mathematical reasoning, but still continue to struggle with high-precision tasks like numerical computation and formal symbolic manipulation. Integrating external tools has emerged as a promising approach to bridge this gap. To overcome these limitations, we propose THOR (Tool-Integrated Hierarchical Optimization via RL). First, we introduce TIRGen, a multi-agent actor-critic-based pipeline for constructing high-quality datasets of tool-integrated reasoning paths, aligning with the policy and generalizing well across diverse models. Second, to perform fine-grained hierarchical optimization, we introduce an RL strategy that jointly optimizes for both episode-level problem solving and step-level code generation. This is motivated by our key insight that the success of an intermediate tool call is a strong predictor of the final answer's correctness. Finally, THOR incorporates a self-correction mechanism that leverages immediate tool feedback to dynamically revise erroneous reasoning paths during inference. Our approach demonstrates strong generalization across diverse models, performing effectively in both reasoning and non-reasoning models. It further achieves state-of-the-art performance for models of a similar scale on multiple mathematical benchmarks, while also delivering consistent improvements on code benchmarks. Our code will be publicly available at https://github.com/JingMog/THOR.